I'm currently researching how current LLMs can affect critical system codebases like Kubernetes or gRPC, and something that I find incredibly interesting is the manner in which they perform these edits, as well as the behavior trends I have noticed. For the most part, models (opus, astra, etc.) uses some set of preconcieved notions about what a codebase of a certain manner should look like, and then edits or creates guardrails for each codebase based on this "idealized" version. This can be helpful in several ways, especially for replicating similar systems. However, this does pose an interesting dilemma about the potential for specification and even more so for specialization. Obviously, after a while, these models are certainly capable of meeting most, if not all demands. However, it definitely takes a ton of compute and time to reach that goal. Maybe a generalized starting point, in the manner that LLMs operate in, isn't the most ideal. Food for thought.